Agent Memory

## Where the vocabulary comes from Agent memory terminology is borrowed from cognitive psychology, which separates the small amount of information a person is holding in mind right now from the much larger store they can recall later, and then splits that long-term store into memory for events, memory for facts, and memory for skills. Agent frameworks adopted the same four labels because they line up with four distinct engineering problems: what to put into the next model call, what history to keep, what durable knowledge to accumulate, and what capability to reuse. The biological analogy is loose — nothing in a language model works like a hippocampus — but the vocabulary is now standard, and an interviewer expects you to use it precisely. ## Working memory Working memory is the state of the task in progress: the user's latest message, the plan, results returned by tools, intermediate notes. Most of it sits in the context window assembled for the current model call, but not all of it needs to. An agent that writes a scratchpad file partway through a long job and re-reads it after each tool result is also using working memory — the file is simply working memory that survives a context rebuild. The defining property is scope, not location: working memory belongs to one task or session and is discarded when that ends. Nothing in it persists unless something deliberately writes it somewhere durable. ## Episodic memory Episodic memory is the record of what happened, in order, with timestamps: on this date the user asked for X, the agent called tool Y, the call failed, the agent retried with different arguments and succeeded. It answers "what did we do last time?" Its natural shape is append-only and time-ordered — you replay or summarize episodes rather than edit them. Raw episodes are verbose, so production systems usually keep a condensed summary per session alongside, or instead of, the full transcript. ## Semantic memory Semantic memory holds durable facts, either distilled from episodes or supplied directly: this customer is on the enterprise plan, the deploy script lives at a particular path, this student struggles with fraction division. The shape is a set of statements rather than a sequence of events. Facts can still expire or be superseded, but they are not inherently tied to the moment they were observed. This is the tier that gets updated and deduplicated, because two different episodes can yield the same fact — or contradictory ones. ## Procedural memory Procedural memory is know-how: a routine the agent has learned or been given and can invoke again. In practice it is stored as executable code, scripts, or structured skill documents rather than as free prose, because a procedure can be tested and a paragraph cannot. The Voyager Minecraft agent is the canonical illustration — it wrote programs, verified them against the environment, kept the ones that worked in a growing skill library, and composed them into harder tasks later. ## Why the split matters The four types have different lifetimes (one task, one session, indefinite, versioned), different access patterns (rebuild every call, replay in order, look up by similarity or key, fetch by name), and different failure modes (overflow, unbounded growth, staleness and contradiction, broken or unsafe skills). Collapsing them into a single vector store of text chunks is the most common design mistake: you end up running similarity search over an event log and retrieving a fragment of an old transcript when what the agent needed was one current fact. ## What the taxonomy does not give you It is a naming scheme, not an architecture. It says nothing about when to write, what to retrieve, or how to decide something has gone stale — those are separate policy questions with their own tradeoffs. Some systems deliberately implement all four in one substrate, such as a directory of files, and let the folder layout carry the distinction. Naming the four types correctly is a warm-up; the interview usually moves quickly to which types your system actually needs and what happens when one of them is wrong. ## Common confusions worth pre-empting Retrieval over a static document corpus is not agent memory in this sense — it is external knowledge the agent reads, not experience it formed and can update. A rolling conversation summary is compressed episodic memory, not semantic memory, until the facts inside it are extracted as standalone statements. And working memory is not "the small memory": it is often the largest thing in play by token count, just the shortest-lived.